The Pentagon has stopped using Anthropic’s AI tools after blacklisting the company, the BBC reported Monday, following a period of tensions over Anthropic’s refusal to provide unrestricted access to its models for certain classified applications. The decision came days after President Trump appointed a senior intelligence official to lead a new AI taskforce — a move framed as bringing military-grade discipline to America’s AI governance. The Anthropic blacklisting is not the first such falling-out between the defence establishment and a leading AI developer: OpenAI’s relationship with the Pentagon has been managed through arm’s-length contractors rather than direct licensing for much of the past two years, and the government’s most recent AI procurement framework explicitly favoured developers willing to remove safety-testing friction for national security use cases. Anthropic’s Claude models, trained under a “Constitutional AI” framework that emphasises refusing harmful requests, are apparently incompatible with some of the Defence Department’s operational requirements — a conflict that was foreseeable from the moment the company’s alignment-first philosophy met the Pentagon’s task-completion-first culture.
The received wisdom
The mainstream technology press has framed this story largely as a corporate-government spat: Anthropic chose its principles and paid a commercial price. The company’s Constitutional AI approach — training models to refuse a defined set of harmful outputs — is described as an admirable if commercially costly choice, with the Pentagon’s departure as evidence that the Trump administration’s hawkish AI posture is crowding out responsible developers in favour of less scrupulous ones. The deeper fear, articulated by AI safety researchers and some Democratic members of the Senate Armed Services Committee, is that the new AI taskforce represents an attempt to militarise American AI development in ways that will produce powerful but ungoverned systems, setting a dangerous precedent that China will inevitably match. OpenAI’s own internal safety firings last week — the company dismissed workers for “mishandling sensitive information” — added to a sense that the AI sector is under pressure to subordinate safety culture to institutional loyalty.
A different read
The mainstream framing contains a real concern buried under a misleading premise, and it is worth separating the two.
The real concern: militarising AI development without robust governance creates systems that are powerful, opaque, and potentially catastrophic if misused. The lesson of every previous dual-use technology — nuclear, chemical, biological — is that the absence of governance frameworks during the developmental window is precisely when they are most needed, and that institutional cultures in military and intelligence organisations do not naturally generate the kind of adversarial safety testing that civilian AI labs perform. The Pentagon’s procurement culture rewards capability demonstration, not safety margin. These are legitimate worries.
The misleading premise: that Anthropic’s Constitutional AI approach is the appropriate baseline for national security AI tools, and that the Pentagon’s frustration with it is therefore simply the triumph of expediency over principle.
This premise fails for several reasons. First, the refusal protocols embedded in models like Claude were designed for consumer and enterprise use cases — preventing the generation of, say, instructions for building explosives in a retail context — not for military targeting or intelligence analysis. A model that refuses to discuss certain kinds of threat information because it was trained on a consumer safety rubric is not safer in a national security context; it is less useful without being meaningfully safer. The Pentagon’s actual risk environment is not “someone using Claude to write malware” but rather “adversaries using AI to accelerate nuclear material identification or biological agent synthesis.” Anthropic’s consumer safety guardrails address neither threat.
Second, the framing misunderstands what the AI taskforce is likely to produce. Trump’s choice of a senior intelligence official to lead the new AI governance body suggests a counterintelligence-oriented approach: primarily concerned with preventing adversarial AI exploitation and maintaining US technical superiority over China and Russia, rather than removing all safety constraints from domestic AI development. Whether that concern is well-executed is a separate question, but presenting it as pure recklessness ignores the genuine threat environment that animates it.
The historical analogy is instructive. In the early Cold War, the US academic-military relationship over nuclear research was deeply uncomfortable in both directions: physicists who had built the bomb were subjected to loyalty investigations, while some of the most important security decisions were made without adequate scientific input. The Oppenheimer clearance revocation was both a security over-reaction and a real example of how the government’s approach to managing dual-use scientists destroyed trust that took decades to rebuild. The AI sector is not there yet, but the trajectory is recognisable. The question is whether the technology companies that refused to engage with defence applications during the last decade — on the grounds that all military AI was ethically compromised — left the field to contractors with weaker safety cultures and less technical credibility. There is a strong argument that they did.
A conservative analysis here is neither triumphalist about the Pentagon’s posture nor naively deferential to Anthropic’s self-presentation. The goal — American technical superiority over adversaries, combined with sufficient internal governance to prevent the worst misuse scenarios — requires both sides to engage honestly with what the other needs. The current trajectory, in which safety-oriented developers are blacklisted and the field is left to less scrupulous actors, serves neither goal.
What to watch
- Who fills the Anthropic gap: Which contractors or AI developers take over the Pentagon’s AI needs? The answer will reveal whether the procurement is moving toward less safety-conscious models or toward specialised national security variants with different but genuine governance frameworks.
- Congressional response: The Senate Armed Services Committee has members from both parties concerned about AI governance. Watch for hearings or framework legislation that might set minimum safety standards even for national security AI procurement.
- Anthropic’s commercial strategy: Whether Anthropic can offset lost government revenue through its enterprise cloud partnerships, or whether the blacklisting triggers a funding squeeze that forces it toward the military market on worse terms.
- The China comparison: Beijing’s AI development for military applications operates under no such friction. The US debate about how much friction is appropriate should be calibrated against what that means for the competitive balance.
— J